A MULTIPLE CLASSIFIER SYSTEM FOR PREDICTING WITH MISSING SENSOR VALUES

Sensors placed in the field frequently suffer faults or communication errors, causing readings to become unavailable for prediction purposes. Statistical approaches are commonly used to generate an artificial value to approximate a missing sensor reading and complete a feature vector for prediction. In this paper, we present a new multiple classifier system that is capable of predicting without artificial approximations of missing values. Each classifier in the system is trained on different feature subsets, so that a classifier exists for each potential combination of available features. We have evaluated the proposed approach on a set of benchmark datasets to establish the generic nature of the problem, and on a real-world environmental event detection application that inspires the need for this work. On benchmark datasets, we artificially vary the percentage of missing values (i) during the prediction phase, and (ii) during both the training and prediction phases. We compare the performance of the prop...